downgrade CUDA toolkit

* make _VSTD workaround

README.md

@@ -44,7 +44,7 @@
   </a>
 
   <a href="https://developer.nvidia.com/cuda-downloads">
-    <img alt="CUDA" src="https://img.shields.io/badge/CUDA-12.6-green?style=flat&logo=nvidia">
+    <img alt="CUDA" src="https://img.shields.io/badge/CUDA-12.4.1-green?style=flat&logo=nvidia">
     </a>
   </a>
 
@@ -182,7 +182,7 @@ See the [example](./example) folder for some examples.
 Prerequisites:
 * `x86_64` Linux instance.
 * Nix with flake support (check out [The Determinate Nix Installer](https://github.com/DeterminateSystems/nix-installer))
-* Nvidia GPU capable of running CUDA 12.6 code.
+* Nvidia GPU capable of running CUDA 12.4.1 code.
 
 From your terminal, run the following command in the root of the source directory to set
 up a build environment. 

nix/cuda.nix

@@ -7,9 +7,8 @@ with pkgs;
 pkgs.stdenvNoCC.mkDerivation {
   name = "cudatoolkit";
   src = fetchurl {
-    url = "https://developer.download.nvidia.com/compute/cuda/12.6.0/local_installers/cuda_12.6.0_560.28.03_linux.run";
-    hash = "sha256-MasEOU5psU3YZW4rRMKHfbGg6Jjf+KdUakxihDgQG5Q=";
-
+    url = "https://developer.download.nvidia.com/compute/cuda/12.4.1/local_installers/cuda_12.4.1_550.54.15_linux.run";
+    sha256 = "367d2299b3a4588ab487a6d27276ca5d9ead6e394904f18bccb9e12433b9c4fb";
   };
   patches = [
     # patch host_defines.h to work with libc++

sxt/algorithm/block/runlength_count.h

@@ -16,6 +16,26 @@
  */
 #pragma once
 
+/*
+ * This is a workaround to define _VSTD before including cub/cub.cuh.
+ * It should be removed when we can upgrade to a newer version of CUDA.
+ *
+ * We need to define _VSTD in order to use the clang version defined in 
+ * clang.nix and the CUDA toolkit version defined in cuda.nix.
+ * 
+ * _VSTD was deprecated and removed from the LLVM truck.
+ * NVIDIA: https://github.com/NVIDIA/cccl/pull/1331
+ * LLVM: https://github.com/llvm/llvm-project/commit/683bc94e1637bd9bacc978f5dc3c79cfc8ff94b9
+ * 
+ * We cannot currently use any CUDA toolkit above 12.4.1 because the Kubernetes 
+ * cluster currently cannot install a driver above 550. 
+ * 
+ * See CUDA toolkit and driver support: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
+ */
+#include <__config>
+#define _VSTD std::_LIBCPP_ABI_NAMESPACE
+_LIBCPP_BEGIN_NAMESPACE_STD _LIBCPP_END_NAMESPACE_STD
+
 #include "cub/cub.cuh"
 #include "sxt/base/macro/cuda_callable.h"
 

sxt/multiexp/base/scalar_array.cc

@@ -18,6 +18,26 @@
 
 #include <vector>
 
+/*
+ * This is a workaround to define _VSTD before including cub/cub.cuh.
+ * It should be removed when we can upgrade to a newer version of CUDA.
+ *
+ * We need to define _VSTD in order to use the clang version defined in 
+ * clang.nix and the CUDA toolkit version defined in cuda.nix.
+ * 
+ * _VSTD was deprecated and removed from the LLVM truck.
+ * NVIDIA: https://github.com/NVIDIA/cccl/pull/1331
+ * LLVM: https://github.com/llvm/llvm-project/commit/683bc94e1637bd9bacc978f5dc3c79cfc8ff94b9
+ * 
+ * We cannot currently use any CUDA toolkit above 12.4.1 because the Kubernetes 
+ * cluster currently cannot install a driver above 550. 
+ * 
+ * See CUDA toolkit and driver support: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
+ */
+#include <__config>
+#define _VSTD std::_LIBCPP_ABI_NAMESPACE
+_LIBCPP_BEGIN_NAMESPACE_STD _LIBCPP_END_NAMESPACE_STD
+
 #include "cub/cub.cuh"
 #include "sxt/base/container/span_utility.h"
 #include "sxt/base/device/memory_utility.h"

sxt/multiexp/bucket_method2/multiproduct_table.cc

@@ -16,6 +16,26 @@
  */
 #include "sxt/multiexp/bucket_method2/multiproduct_table.h"
 
+/*
+ * This is a workaround to define _VSTD before including cub/cub.cuh.
+ * It should be removed when we can upgrade to a newer version of CUDA.
+ *
+ * We need to define _VSTD in order to use the clang version defined in 
+ * clang.nix and the CUDA toolkit version defined in cuda.nix.
+ * 
+ * _VSTD was deprecated and removed from the LLVM truck.
+ * NVIDIA: https://github.com/NVIDIA/cccl/pull/1331
+ * LLVM: https://github.com/llvm/llvm-project/commit/683bc94e1637bd9bacc978f5dc3c79cfc8ff94b9
+ * 
+ * We cannot currently use any CUDA toolkit above 12.4.1 because the Kubernetes 
+ * cluster currently cannot install a driver above 550. 
+ * 
+ * See CUDA toolkit and driver support: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
+ */
+#include <__config>
+#define _VSTD std::_LIBCPP_ABI_NAMESPACE
+_LIBCPP_BEGIN_NAMESPACE_STD _LIBCPP_END_NAMESPACE_STD
+
 #include "cub/cub.cuh"
 #include "sxt/algorithm/iteration/for_each.h"
 #include "sxt/base/device/memory_utility.h"

sxt/multiexp/bucket_method2/multiproduct_table_kernel.h

@@ -18,6 +18,26 @@
 
 #include <utility>
 
+/*
+ * This is a workaround to define _VSTD before including cub/cub.cuh.
+ * It should be removed when we can upgrade to a newer version of CUDA.
+ *
+ * We need to define _VSTD in order to use the clang version defined in 
+ * clang.nix and the CUDA toolkit version defined in cuda.nix.
+ * 
+ * _VSTD was deprecated and removed from the LLVM truck.
+ * NVIDIA: https://github.com/NVIDIA/cccl/pull/1331
+ * LLVM: https://github.com/llvm/llvm-project/commit/683bc94e1637bd9bacc978f5dc3c79cfc8ff94b9
+ * 
+ * We cannot currently use any CUDA toolkit above 12.4.1 because the Kubernetes 
+ * cluster currently cannot install a driver above 550. 
+ * 
+ * See CUDA toolkit and driver support: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
+ */
+#include <__config>
+#define _VSTD std::_LIBCPP_ABI_NAMESPACE
+_LIBCPP_BEGIN_NAMESPACE_STD _LIBCPP_END_NAMESPACE_STD
+
 #include "cub/cub.cuh"
 #include "sxt/algorithm/block/runlength_count.h"
 #include "sxt/base/device/stream.h"